基于边缘组合算法的似物性采样研究

摘要

This paper uses edge information to do qualitative analysis of object window and designs a window sampling strategy through edge grouping,which reduces the proposal number and ensures the precision of windows.Then it sorts these windows according to an object estimate based on edges and provides object proposals with high-accuracy for object detection in a short time.This algorithm shows better balance between accurate and computational efficiency in the VOC2007 database.When 1 000 proposals are given each picture,the proposed method runs in 0.2 s and achieves over 79％ recall at the more challenging overlap of 0.7.Comparing with other object proposal methods,ours has more universal and applicability.%利用图像中的边缘信息对似物性窗口进行定性分析,设计了一种采用边缘组合算法的采样,降低了采样数量的同时保证了窗口精度,并用一种基于边缘的似物性估计为这些窗口排序,在较短的时间内为物体检测算法提供高精度的建议窗口.在VOC2007测试集中该算法在性能上达到了真正的平衡,限定每张图片输出1 000个窗口时,不仅达到了0.2s的计算速度,而且在0.7的重叠率条件下命中率高达79％,与其他似物性采样相比,该算法具有更强的通用性和适用性.